A Federated Learning Protocol for Spiking Neural Membrane Systems.

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Machine learning

Background:

• Deep learning models, while effective, differ significantly from biological brain function in energy consumption, communication, and learning.
• Spiking neural networks (SNNs) aim to bridge this gap by mimicking biological neuron behavior.
• Layered Spiking Neural P systems (LSN P systems) are a type of SNN designed for classification tasks.

Purpose of the Study:

• To investigate the application of LSN P systems within a federated learning (FL) framework.
• To analyze the privacy risks associated with pre-trained LSN P systems using membership inference attacks.
• To evaluate the performance and efficiency of LSN P systems trained in an FL environment.

Main Methods:

• Implementing LSN P systems in a client-server FL architecture with horizontally partitioned data.
• Conducting membership inference attacks to assess privacy vulnerabilities of pre-trained models.
• Performing comparative experiments to evaluate LSN P system performance against other federated algorithms.

Main Results:

• LSN P systems trained via federated learning exhibit enhanced accuracy.
• These systems demonstrate faster convergence rates compared to traditional federated algorithms.
• Privacy analysis revealed potential vulnerabilities through membership inference attacks.

Conclusions:

• LSN P systems present a promising, biologically plausible approach for federated learning.
• They offer performance advantages in terms of accuracy and training speed over existing methods.
• Further research is needed to address privacy concerns in federated LSN P systems.